Advancing Bayesian Optimization via Learning Correlated Latent Space
Authors: Seunghun Lee, Jaewon Chu, Sihyeon Kim, Juyeon Ko, Hyunwoo J. Kim
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate Co BO to nine tasks on a discrete space in three different Bayesian optimization benchmarks, which consist of arithmetic expression fitting tasks, molecule design tasks named dopamine receptor D3 (DRD3) in Therapeutics Data Commons (TDC) [26] and Guacamol benchmarks [27]. The arithmetic expression task is generating polynomial expressions that are close to specific target expressions (e.g., 1/3+x+sin(x x)) [10, 11, 13, 14, 28], we set the number of initialization points |D0| to 40k, and max oracle calls to 500. In Guacamol benchmark, we select seven challenging tasks to achieve high objective value, Median molecules 2, Zaleplon MPO, Perindopril MPO, Osimertinib MPO, Ranolazine MPO, Aripiprazole similarity, and Valsartan SMART. The results for the last three tasks are in the supplement. The goal of each task is to find molecules that have the most required properties. For every task of Guacamol benchmark, we set the number of initialization points to 10k, and max oracle calls to 70k. DRD3 task in the TDC benchmark aims to find molecules with the largest docking scores to a target protein. In DRD3, the number of initialization points is set to 100, and the number of oracle calls is set to 3k. |
| Researcher Affiliation | Academia | Seunghun Lee , Jaewon Chu , Sihyeon Kim , Juyeon Ko, Hyunwoo J. Kim Computer Science & Engineering Korea University {llsshh319, allonsy07, sh_bs15, juyon98, hyunwoojkim}@korea.ac.kr |
| Pseudocode | Yes | Algorithm 1 Correlated Bayesian Optimization (Co BO) |
| Open Source Code | No | The paper does not explicitly state that source code for the described methodology is provided or publicly available. |
| Open Datasets | Yes | We evaluate Co BO to nine tasks on a discrete space in three different Bayesian optimization benchmarks, which consist of arithmetic expression fitting tasks, molecule design tasks named dopamine receptor D3 (DRD3) in Therapeutics Data Commons (TDC) [26] and Guacamol benchmarks [27]. |
| Dataset Splits | No | The paper mentions 'number of initialization points' and 'max oracle calls' for the benchmarks, and 'training data for the VAE model and surrogate model', but does not provide specific details on train/validation/test splits (e.g., percentages, sample counts, or explicit references to standard splits for reproducibility). |
| Hardware Specification | No | The paper does not provide any specific hardware details such as GPU or CPU models, memory specifications, or cloud/cluster configurations used for running the experiments. |
| Software Dependencies | No | The paper mentions various software components and models such as 'Gaussian process (GP)', 'VAE', 'Thompson sampling', 'sparse GP', 'RBF kernel', 'deep kernel learning (DKL)', 'SELFIES VAE', and 'Grammar VAE', but it does not provide specific version numbers for any of these or other ancillary software dependencies. |
| Experiment Setup | Yes | The arithmetic expression task... we set the number of initialization points |D0| to 40k, and max oracle calls to 500. For every task of Guacamol benchmark, we set the number of initialization points to 10k, and max oracle calls to 70k. In DRD3, the number of initialization points is set to 100, and the number of oracle calls is set to 3k. ...we retrain a latent space after Nfail accumulated failure of updating the optimal objective value. ...batch size Nb. ...We set the Lipschitz constant L as the median of all possible gradients of slopes. ...We set c as the expected Euclidean norm between two standard normal distributions... |